Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Learning Foundry
Inspiration
AI can increase production velocity faster than people can build and maintain the mental models needed to participate. The problem is not simply remembering more. Human confidence, source-backed knowledge, agent synthesis, and validated behavior are different epistemic states, yet most tools collapse them into chat history, a completion mark, or a mastery score.
Learning Foundry makes those differences inspectable. It creates an environment where a person and an agent learn from the same approved evidence while retaining distinct learning states.
What it does
Learning Foundry accepts local and approved online sources through one provenance-preserving pipeline. It synthesizes them into a revisable Living Theory of concepts, claims, relationships, assumptions, boundaries, decisions, contradictions, and open questions.
The same theory generates source-grounded explainers, varied understanding checks, transfer tasks, and constrained interactive micro-worlds for the human. It also grounds versioned Codex capabilities with explicit boundaries, evaluations, failures, and approval gates.
Practical results return as append-only evidence. A correction can create targeted human review, a theory revision proposal, and a structured capability revision request without overwriting the original source, silently changing memory, or activating a skill.
How we built it
The desktop application uses Electron, React, TypeScript, Vite, Zod, and an append-only JSONL evidence ledger. Canonical events are replayed into separate human-memory, agent-memory, shared-theory, understanding-gap, source, learning, and capability projections. Every generated artifact and consequential decision retains provenance.
The prepared design-density journey is committed sample data and runs offline. A constrained micro-world lets the learner predict, manipulate spacing, hierarchy, and visible information, then record an observation and reflection as separate evidence. A prepared execution adapter keeps the demonstration deterministic; an optional live Codex adapter requires explicit consent and records its bounded prompt context and fallback behavior.
How we used Codex and GPT-5.6
We built the project collaboratively in Codex with GPT-5.6. Codex helped turn the initial product thesis into a sequenced issue map, implement each vertical slice, maintain architecture decisions, and run the verification loop. Pull-request findings became focused regression tests and fixes covering atomic ledger transitions, provenance, network and IPC boundaries, accessibility, state retention, deterministic replay, approval gates, and prepared-demo identity.
Human judgment remained responsible for product direction, interaction design, epistemic boundaries, review decisions, and what entered the canonical evidence model. This collaboration is also the product thesis: agents can accelerate implementation while the environment keeps rationale, evidence, uncertainty, and participation visible to the human.
Challenges
The hardest problem was preserving meaning across the complete loop. It was easy to make each screen work independently; it was harder to ensure that a prediction, interaction, reflection, capability execution, and correction remained distinct, traceable events whose projections could be replayed deterministically.
We also had to resist false precision. Learning Foundry exposes qualitative, evidence-backed understanding gaps, but it does not claim to measure cognition or solve cognitive debt.
What we learned
Shared context is not shared understanding. A useful human-agent system needs stable identifiers, inspectable provenance, explicit epistemic states, and practical evidence that can challenge both learning material and agent behavior. Dynamic learning media is valuable when it reveals a causal relationship, not when it merely adds interaction.
What's next
The current design-density workspace proves a domain-independent contract. Next steps include additional bounded micro-world renderers, richer source extractors, user-authored consolidation policies, and generated tools or plugins that use the same evaluation and approval lifecycle.
Log in or sign up for Devpost to join the conversation.